Overview
Embedding models transform raw data (text, images, audio) into dense numerical vectors that capture semantic meaning, enabling AI applications to understand and compare content.
How They Work
- Input: Raw data (text, image, etc.)
- Encode: Neural network processes input
- Output: Fixed-size vector (e.g., 768 dimensions)
- Property: Similar inputs → similar vectors
Types
Text Embeddings
- Sentence-BERT
- BGE models
- OpenAI text-embedding-*
- Cohere Embed
Multimodal
- CLIP (text + image)
- Gemini Embedding 2
- UForm
Specialized
- Code embeddings
- Audio embeddings
- Graph embeddings
Key Characteristics
Dimensionality
- Small: 384-512 (fast, less info)
- Medium: 768-1024 (balanced)
- Large: 1536-4096 (rich, slower)
Context Window
- Max input length
- Typically 512-8192 tokens
- Affects chunk size decisions
Choosing Models
Considerations
- Task: Search, classification, clustering
- Language: Monolingual vs multilingual
- Domain: General vs specialized
- Performance: Speed vs quality
- Cost: Self-hosted vs API
Evaluation
- MTEB Leaderboard
- Task-specific benchmarks
- Domain testing
Popular Models (2026)
- NV-Embed (NVIDIA)
- BGE-M3 (BAAI)
- Gemini Embedding 2
- text-embedding-3-* (OpenAI)
- Voyage embeddings
- Jina Embeddings v4/v5
Pricing
- Open Source: Free (hosting costs)
- APIs: Usage-based pricing
- Self-hosted: Compute infrastructure
